Development and Validation of a Convolutional Neural Network Framework Based on Ultrasound Imaging for Multi-Classification of Superficial Soft Tissue Masses
In brief
AI ultrasound tool classifies malignant soft-tissue masses with 98% accuracy, beats senior radiologists
In a multi-center study of 3,168 patients, the ST-USNet deep-learning system distinguished malignant from benign superficial soft-tissue lesions with an area under the curve of 0.984 and outperformed senior radiologists in a reader test. The model also accurately subtyped sarcoma, lymphoma and metastases, but prospective validation is still needed.
- Journal
- Academic radiology (Q1)
- Published
- 7 September 2026
- Study design
- Prospective / inception cohort
- Evidence level
- Level 2, Moderate (CEBM 2b)
- Authors
- Mengjie Wu, Yulu Wu, Ao Li, Yilei Shi, Lehang Guo, Chunlei Li, et al.
- PMID
- 42705922
- DOI
- 10.1016/j.acra.2026.08.092
Why clinicians should know about it
- Picked for Radiology, Radiation Oncology, Nuclear Medicine, Medical Physics and Imaging (paper of the day, 10 September 2026): CNN multi‑classification of superficial soft‑tissue masses, high AUCs
Abstract
RATIONALE AND OBJECTIVES: Superficial soft tissue masses (STMs) represent a diagnostic dilemma in clinical practice, with ultrasound (US) being the front-line imaging modality available globally. However, the high complexity of STMs in imaging makes subjective evaluation highly dependent on experience, frequently causing inconsistent malignancy assessments. This inconsistency triggers unnecessary benign biopsies and delays treatment for malignant STMs. We developed ST-USNet, a multitask convolutional neural network framework to classify superficial STMs based on manually drawn regions of interest. MATERIALS AND METHODS: This retrospective study included US images of 3168 patients (median age, 58 years; IQR, 46-68 years) with STMs from four institutions between March 2015 and October 2024. The ST-USNet was developed and validated on multi-center data. Its performance was then evaluated on a separate, independent test cohort. The diagnostic performance of ST-USNet was compared with that of radiologists using McNemar tests. RESULTS: The ST-USNet was composed of four sub-models (SM-1, SM-2, SM-a, and SM-b). In the validation cohort, SM-1 was trained to distinguish malignant from benign STMs (AUC: 0.984); SM-2 was to classify the malignant STM subtypes including sarcoma, lymphoma, and metastatic carcinoma (AUC: 0.932, 0.909, 0.922); SM-a was to discriminate between aggressive and indolent lymphoma (AUC: 0.951). SM-b was designed to explore the identification of metastatic carcinoma origin (thyroid, breast, respiratory, digestive, and reproductive systems) as a preliminary analysis; however, due to limited sample sizes, these results should be interpreted as exploratory. ST-USNet achieved high AUCs on the validation cohort and remained effective, albeit with slightly lower performance, on an independent test cohort. In a preliminary reader study (4 radiologists, 85 cases), ST-USNet either significantly outperformed senior radiologists (p = 0.012) or performed comparably to them (p = 0.267, 0.092, 0.332) in all classification tasks and effectively enhanced diagnostic accuracy for both junior and senior radiologists when used as an assistive tool. CONCLUSION: ST-USNet serves as an effective and practical decision support system for superficial STMs classification in clinical oncology, though multi-center prospective validation and continuous model updating are required before clinical deployment.
Abstract as published, via PubMed.
For healthcare professionals. The summary is generated by AI from the published abstract, and the evidence level is assigned automatically from the study design on the Oxford CEBM hierarchy. Neither is medical advice. Read the full paper before changing practice.